Video summary
I Turned Claude Into the Ultimate Second Brain
Main summary
Key takeaways
Claude for a “Second Brain / AI OS”
- The creator uses Claude (referred to as “Claude Fable” in subtitles) as the core of an AI “second brain” system.
- The system is framed as an always-aware cofounder/teammate that can:
- Understand both business and personal context
- Automate workflows based on that context
Model release: availability and pricing guidance
Release timing
- Time-limited availability: June 9 → June 22 (on a subscription)
- After that window, access appears to switch to usage credits
Cost warning (relative pricing)
- Described as ~2× more expensive than Opus
- Example token rates mentioned:
- $10 / 1M input tokens
- $50 / 1M output tokens
- Practical guidance:
- Experiment during the 2-week window
- Expect it to consume limits faster
Cyber guard rails vs flagship comparison
- The model is positioned as similar to Anthropic’s flagship “Claude” tease (mentioned as “Mythos”)
- But with more “cyber guard rails”
AI Operating System architecture framework: “The Four C’s”
The framework is taught as two layers:
- Second Brain = Context + Connections
- AI OS = Capabilities + Cadence
C1: Context
- Represents who you are and what your business is
- Implemented as a routing tree that points agents to:
- The right files
- Rules
- References
- Skills
- Wikis
- Etc.
Context management (“pulse checks”)
- If an agent can’t find something quickly, you may need:
- Architecture changes
- File splitting
- Manual drill-through should feel intuitive, and the agent should behave similarly.
C2: Connections
- Represents live / updated data sources
- Distinguishes:
- Static knowledge (docs/transcripts stored in files)
- Dynamic/live systems (email, ClickUp, QuickBooks P&L)
- Connections are typically integrated via:
- APIs / API endpoints / CLIs
- Wired through a tool harness
C3: Capabilities
Capabilities cover skills/agents/automations.
Skills
- Can be either:
- Complex workflows
- Or reusable prompts
- Skills can be broken into specialized “assembly line” stages (one AI task per stage).
Iterative improvement loop
- Every skill use generates feedback
- The creator then updates the skill
- Emphasis: no “finished product” mindset
Delegation and guardrail tuning
- Many custom skills (and “sub-agents”) are used inside the coding/workspace tool (referenced as “Claude code” features).
- The approach encourages delegation to cheaper models (e.g., Sonnet/Haiku) for parallel tasks, then consolidating results.
- Guardrails may be overly trigger-happy and require tuning over time.
C4: Cadence
Cadence covers automations that run over time, and requires earned trust.
Key tradeoffs
- As autonomy increases, you get more:
- Cost
- Risk
- Maintenance
- Even in production, automations need:
- Visibility / ownership
- Periodic checking
Trigger types
- Manual
- Event-based (e.g., new email / customer booking)
- Scheduled (weekly routines)
Deployment styles
- Can include different execution approaches (routines/loops/deterministic scripts)
- Mentions external deployment tools conceptually
Security / permission layer concept
A prompt is not a permission layer.
- Agents should operate with keys that have the minimum required access
- Example: read-only transcript access
- Safety lesson mentioned:
- An agent allegedly sent emails with a wrong discount code due to overly proactive task interpretation
Concrete setup details: how their AI OS is organized
“Herk 2” OS/project
- The system is organized as “Herk 2” (as shown in the subtitles/labels).
Main components
- A main project containing:
- The routing tree
- Rules, references, skills
- Memory files
- A knowledge base folder/path layout, including:
- Wiki path
- Hot cache
- Master index
- Navigation method
Sub-agents and skills as the primary productivity unit
- “Sub-agents” and “skills” are presented as the #1 productivity feature.
“Other Worlds” folder
- A folder that houses other frequently used code projects moved into the main OS project.
- Motivation:
- Easier syncing to GitHub
- Better context for the OS to operate across repos
Token/context management observations
- Example shows a “/context” starting around ~40k tokens, mostly system tools.
- Large folder projects can still be manageable when architecture is organized.
Examples of Claude Fable outputs
- A one-shot “/goal” prompt to generate a journey/about style YouTube video using stored context.
- A noted error:
- Static data caused outdated numbers (e.g., YouTube subscriber count)
- Suggested fix: use live connections so numbers refresh dynamically
- Another one-shot output:
- Builds an interactive relationship map / very clean interface connecting concepts, tools, and where they appear in transcripts/videos
- Usage/time expectations:
- Mentions session depletion and runtime expectations
- Example: a heavy task taking roughly ~21 minutes
Usage tips (tutorial-like recommendations)
- Treat the model as a thought partner
- Brainstorm + devil’s advocate
- Don’t treat it as unquestioned authority
- Interview yourself
- Use a skill like “grill me” to extract deep knowledge into the system via many questions
- Verify the work
- Use dynamic workflows + visual/manual testing (e.g., “playwright-style clicking”)
- Iterate toward roughly ~92% correctness faster, then refine
Tool-agnostic mindset
- The creator stresses the “real IP” is:
- Folders/files
- Skills
- Routing logic
- The system should be portable:
- They claim you can switch between model providers/tooling (e.g., Claude vs others referenced in subtitles)
- Without rebuilding the whole system
FAQ-like points
- Cost
- Based on plan/session limits (mentions a $200/month plan and rarely hitting limits)
- Data handling
- If using Claude models, data goes to Anthropic (closed source), so sensitive data concerns may apply
- Coding skills
- Not required for day-one setup; use their GitHub repo / course
- If the model is confident but wrong
- Update cloudmd + skills
- Treat mistakes as system-improvement data
- Live connections
- Done with API keys/endpoints and scoped keys for least privilege
- Team adoption
- Team members should learn first
- Shared knowledge should live in centralized locations (e.g., ClickUp / Slack / Notion / Drive) with read-only access
- Biggest risk
- Adoption: shared knowledge must stay up to date
Main speakers / sources (as referenced)
Speaker / creator
- Nate Herk
- The narrator describes it as “Nate Herk’s executive assistant”
- References “my OS” and “my business” as context
Model / product sources
- Anthropic
- Claude (including “Claude Mythos” reference)
Community / industry voices
- Andrej Karpathy
- Boris Churnney (spelling uncertain)
- Matt PCO
- Credited for the “grill me” skill idea